Municipal　solid　waste　(MSW)　has　characteristics　such　as　large　component　differences　and　unstable　calorific　value.　Flame　combustion　automatic　state　identification　is　beneficial　for　achieving　stable　control　of　the　MSW　incineration　(MSWI)　process　in　terms　of　feedback　control.　The　industrial　site　usually　recognizes　the　combustion　state　based　on　the　visual　perception　mechanism　of　domain　experts＇　experience.　Based　on　the　widely　application　of　convolutional　neural　network　(CNN)　in　the　field　of　image　recognition,　this　article　chooses　Lenet-5　as　the　recognition　model　of　flame　combustion　state　recognition.　In　order　to　obtain　an　appropriate　CNN　model　to　automatically　extract　flame　image　features,　the　influence　of　multiple　CNN　networks　on　flame　image　feature　is　designed　and　tested.　The　optimal　structure　setting　of　the　CNN　network　is　obtained.　This　study　can　provide　support　for　practical　applications　in　industrial　sites.　©　2023　IEEE.